What good failure evidence looks like

Triage time is the metric nobody tracks and everybody feels.

The five required artefacts

• Full transcript with turn numbers.
• Tool trace with arguments, results and timing.
• Expected vs actual, rendered as a diff.
• Per-evaluator verdicts with reasoning.
• A seed that reproduces the run exactly.

Point at the turn

The report should name the exact turn where behaviour diverged. Scrolling a 40-turn transcript looking for the moment is where triage time goes.

One click to permanence

From the evidence view, promoting the case into the regression suite should take one action and keep the seed attached.

What to take from this

• Evidence, not scores, is what gets bugs fixed.
• Name the diverging turn explicitly.
• Keep the repro seed with the case forever.